Skip to content

[Draft] V2:隔离式 live acceptance、fault probes 与兼容性证据(stacked) - #555

Draft
tbrandenburg wants to merge 31 commits into
Intelligent-Internet:mainfrom
tbrandenburg:upstream/v2-exclusive-acceptance
Draft

tbrandenburg wants to merge 31 commits into
Intelligent-Internet:mainfrom
tbrandenburg:upstream/v2-exclusive-acceptance

Conversation

@tbrandenburg

Copy link
Copy Markdown

范围与依赖

关联 #550;验证跟踪:tbrandenburg#6 。

Stacked / 非独立 diff:依赖 #551 → #552 → #553 → #554;当前 base 为 main。 本层主要是 reproducible acceptance / fault probes、hardened default V2 session smoke、V1 smoke 证据检查和独立 QA 记录。聚焦本层比较:
tbrandenburg/opencode-a2a@upstream/v2-exclusive-integration...upstream/v2-exclusive-acceptance

可复现验收

  • Docker 内 loopback-only upstream,无 published upstream port;先 authenticated-ready,再证明 host 与独立 peer 不能访问。Private HOME/XDG/workspace、随机 separate credentials、精确 container/PID teardown。
  • V2 2.0.28 + 实际观察的 openai/gpt-6-luna-fast,真实 REST/JSON-RPC stream/unary、follow-up、parallel / same-session serialization、local queued / active cancel、read allow/reject、string/multiselect form 与 owner/typed/stale callback checks。
  • 额外证明 session hijack 与 outside-root/changed-directory metadata 不产生 native admission,真实 external-directory read 被 policy 拒绝且无文件内容泄漏。
  • Positive run 14 native admissions / 1 global SSE。输出只从 assistant artifacts 按 append/snapshot 规则重建,不能被用户输入回显满足。
  • 六个 fresh-pair negative scenarios:mid-run SSE gap、lost terminal、lost admission response、真实 provider 拒绝 synthetic invalid credential、input-required 时断线、callback expiry。每次一个 admission、不误报成功、不 blind resend。
  • V1 smoke 改用随机 credential/private curl config,避免 argv 明文;sentinel 只检查 assistant artifact。独立 V1 1.18.19 expanded matrix 12/12,permission allow 证明精确文件 side effect、reject 证明 native rejected artifact/无文件;6/6 cancel 后额外 46 秒无 terminal overwrite。

Test-only fault proxy/control endpoint 不是产品 API,不部署到公开服务。没有 mock model/provider positive runs;故障只注入传输,provider failure 单独使用 synthetic invalid credential。

验证与证据边界

最终 acceptance head dfa15f7:bash ./scripts/doctor.sh 1021 passed,0 failed,92.68% coverage;secret scan、lint/format、mypy、critical coverage、build/wheel smoke 通过。Conformance 8/8;lead 又在该精确 checkout 运行默认 V2 session smoke,通过认证拒绝与 identity/location/session routes,未提交 model prompt。生产 95 文件、6 个 smoke/harness 文件、acceptance JSON 与 author da51304 比较零差异;frozen runtime lock 与 build constraints 也一致。

Exact-head regression结果见 docs/evidence/v2-exclusive-upstream-validation.md;author真实模型证据见 docs/evidence/v2-exclusive-acceptance.json。二者分开记录,不能把 shared-source live evidence 冒充每个 PR checkout 的 live rerun。生产源码与经过 live 验证的 author source tree 逐文件一致,公开 SHA-256 用分组形式记录;secret scanner 仍启用,合成 test credentials 只有逐项窄 allowlist。

最终 pinned/head checks:doctor.sh、conformance 8/8、默认 V2 session smoke、V1 live smoke 及新 V2 positive/negative probes。Live reasoning 未验证;一般 shared-session、automatic replay/restart continuation、unary interventions、rich forms、attachments、request-scoped models、arbitrary tools 与 a2a_call 不在受限支持范围。

结果是 restricted profile 实现与其 scoped acceptance,不是 upstream 已发布完整 V2。原 epic / #4/#5/#6 不因该 narrowed profile 自动关闭;fork 非独占后续: tbrandenburg#9 。请 maintainer 明确接受范围并按依赖顺序审查,保持 Draft,不自动 merge。

Tom Brandenburg added 30 commits October 25, 2026 15:14
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant